Essay

The Big Hook

Why Business Becomes Code—and Why Relationship Becomes the Durable Advantage

· Consentful Cybernetics
business-as-codeartificial-intelligencerelational-capabilityconsentful-cyberneticscooperationsemantic-infrastructureorganizational-design

An employee has a question. They could search the company manual, call a coworker, write to a manager, or reconstruct the answer from old messages. Increasingly, they ask an artificial intelligence.

At first, the AI helps them write. Soon, it helps them interpret. Then it helps them decide. Eventually, it helps them act.

This is how the change enters the organization: not as a declaration of institutional transformation, but as an ordinary convenience. Someone wants to get something done. A public cognitive utility is waiting in the browser. The employee asks it to summarize a contract, prepare a proposal, explain a regulation, classify a customer complaint, design a workflow, compare vendors, write code, or recommend what should happen next.

Consider the question another way: Would you ask your employees to conduct company business from a public-library computer?

Probably not. Yet many organizations now allow some portion of their institutional language, judgment, memory, and problem-solving to pass through cognitive infrastructure they neither own nor meaningfully govern. The immediate concern is usually confidentiality. That concern is real, but it is not the deepest change. The deeper change is that the organization has begun transferring part of its operative cognition into an external machine.

The first hook is convenience. The second is dependence. The big hook arrives when machine-mediated participation becomes so normal that a business must make itself legible to machines simply to remain available for ordinary exchange.

Regular people use AI. Businesses moderate that use. The employee becomes hybrid. AIs begin using humans to communicate, while humans use AIs to communicate. Standards emerge for hybrid identity and trust. Businesses adopt deeper automation. Then, gradually and unevenly, businesses become code—not because people disappear, but because people are no longer the only systems carrying the business.

Every Outed Skill Begins Becoming Infrastructure

“Inevitable” is a dangerous word. It can conceal assumptions, flatten uncertainty, and turn a direction into a prophecy. Here it means something narrower:

Wherever competition can act upon a capability that has become sufficiently legible to reproduce, continuing pressure will exist to encode, distribute, purchase, imitate, or automate that capability.

A skill has been outed when its productive regularities have become observable enough to infer. It may be outed through explicit instruction, but it may also be demonstrated through thousands of completed transactions. It may become visible through corrections, decisions, outcomes, rankings, customer reactions, or the accumulated traces of ordinary work.

The system does not need to understand the practitioner as the practitioner understands themself. It does not need to reproduce every reason, sensation, hesitation, or act of judgment. It only needs to reproduce enough of the useful result to reduce the scarcity premium attached to the original capability.

This is not new. The history of computerization is not primarily a history of machines swallowing occupations whole. It is a history of work being decomposed into tasks, with technology substituting for activities that can be sufficiently specified while complementing others. Autor, Levy, and Murnane’s task-based account remains foundational: automation proceeds unevenly through occupations, following the parts of work that can be rendered routine, rule-governed, or otherwise machine-legible.

Generative AI expands the surface of what can be reproduced because it does not always require a programmer to specify the procedure directly. In a large deployment involving more than five thousand customer-support agents, AI assistance improved productivity most among less-experienced and lower-performing workers. The findings were consistent with the system disseminating practices associated with stronger workers across the workforce. Controlled research on professional writing has likewise found that generative AI can reduce completion time and improve measured output quality.

The important movement is not simply from human work to machine replacement. It is from unevenly distributed capability, to observable pattern, to encoded assistance, to widely distributed capacity, and finally to reduced scarcity of the original skill.

Once one company achieves that compression, comparable organizations face pressure to respond. They may adopt the same capability, acquire an equivalent, specialize above it, or leave the market. What was once an advantage becomes an expectation. What was once expertise becomes a feature. What was once a feature becomes infrastructure.

A secret may remain a moat. An expressed capability has already begun becoming infrastructure.

This does not mean every attempt succeeds. Human-machine combinations are not automatically superior to either humans or machines alone. A meta-analysis covering more than one hundred experiments found substantial evidence of AI augmenting human performance, but no average synergy against the stronger of the human-only or AI-only alternatives. Combined systems often performed worse than whichever component was already best.

The frontier is jagged. Some capabilities yield quickly. Some resist. Some appear to be encoded until changing circumstances reveal that the model captured a surface correlation rather than the operative competence. Some forms of machine assistance make an average performer better while distracting an expert. Some improve individual output while compressing the diversity of outputs across a population.

None of this removes the pressure. It means the conversion happens capability by capability rather than occupation by occupation, and organization by organization rather than all at once.

The system does not need to reproduce the expert’s soul. It only needs to make the expert’s scarcity less economically decisive.

The Hybrid Employee Is a Transitional Form

We commonly describe the employee as “using AI.” That language understates the change.

The operative unit is becoming a compound of person, model, instructions, retrieved context, tools, permissions, institutional process, and human review. This hybrid can read more material, generate more alternatives, retain more working context, and operate across more domains than the unaided individual. It is neither simply an employee nor simply a software application. It is a hybrid institutional actor.

This immediately raises an uncomfortable question: Whose action is it?

Suppose an employee asks a model to interpret a policy, retrieve relevant records, compare options, draft a response, and initiate an approved transaction. Was the result produced by the employee’s judgment, the company’s policy, the model’s learned regularities, the vendor’s hidden system instructions, the retrieved documents, the person who wrote the prompt, the person who approved the workflow, or the delegated credential through which the action occurred?

The answer is not merely philosophical. It determines authority, responsibility, evidentiary value, accountability, and recourse.

The hybrid employee cannot be governed adequately as though the AI were a calculator. Nor can the hybrid be treated as an autonomous agent detached from the person and institution whose standing it carries.

For a while, the human serves as an adapter. One employee asks an AI to compose a message. Another receives the message and asks a different AI to summarize it, evaluate it, compare it with policy, and prepare a response. The visible path is human to human, but the functional path is already closer to organization, human, agent, human, agent, organization.

The humans may continue to approve, contextualize, redirect, and bear responsibility. Yet an increasing portion of the compression and interpretation is occurring elsewhere. Eventually, repeatedly translating through a human-shaped interface becomes inefficient. Agents begin exchanging structured state more directly.

The hybrid employee is therefore unlikely to be the endpoint. It is a temporary adapter between human-shaped businesses and machine-shaped coordination.

Software Is Changing Again

In 2017, Andrej Karpathy proposed a distinction between Software 1.0 and Software 2.0. Software 1.0 is the familiar form: explicit instructions written by human programmers. Software 2.0 is learned. Instead of specifying each operative instruction, developers specify architectures, training data, and evaluation criteria, allowing optimization to produce the parameters that govern behavior.

Karpathy argued that any domain in which desirable performance can be evaluated repeatedly, but the program itself is difficult to write explicitly, becomes a candidate for this transition. In his 2025 Y Combinator talk, “Software Is Changing (Again)”, he extended the taxonomy to Software 3.0: large language models functioning as a new kind of computer, with natural language increasingly serving as a programming interface.

The progression is useful because it reveals a change in where human intention enters the machine. In Software 1.0, a person writes instructions. In Software 2.0, people define examples, objectives, data, and evaluation surfaces from which behavior is learned. In Software 3.0, people increasingly express intent in language, and a model interprets that intent within a surrounding harness of context, tools, memory, and permissions.

Business requires another movement.

A language-directed agent can generate an answer or pursue an objective. It cannot legitimately participate in institutional life merely because it is capable of doing so. Once agents cross organizational boundaries, capability is not enough. The participating systems must know which entity is speaking, whom the agent represents, which authority has been delegated, what information may be used or revealed, which commitments the agent may make, what requires direct human judgment, which definitions and policy versions govern, what evidence must survive, how permission may be withdrawn, what happens when the interaction fails, and who remains responsible for repair.

This is the movement beyond language as programming. The program must carry the institutional conditions under which language may become action.

We might call that next layer consentful software, constitutional software, or business as code. The label matters less than the shift:

The machine must represent not only what can happen, but what may happen, for whom, under which state, with what consequence, and according to whose continuing permission.

The Business Was Always a Program

Business as code sounds futuristic only because we are accustomed to confusing the business with its people, buildings, products, and legal shell.

A business has always operated through program-like structures: roles, permissions, triggers, conditions, thresholds, approvals, state changes, escalation paths, memories, sanctions, exceptions, and outputs. A purchase order changes institutional state. A signature activates authority. A delivery creates an obligation. A threshold triggers review. A missed deadline initiates escalation. A representation creates reliance. A breach opens a path toward remedy.

None of that began with artificial intelligence. The difference is that much of this program has historically been carried inside people—in habits, professional judgment, stories, meetings, organizational memory, informal relationships, and the ability to notice that a formal rule should not be applied mechanically in a particular case.

The organization was always executable. Humans were simply its primary runtime.

Ronald Coase’s theory of the firm helps explain why this runtime took the form of an organization. Firms internalize activities partly because arranging every action through separate market transactions would impose recurring costs of discovery, negotiation, contracting, supervision, and coordination. As those costs change, the boundary of the firm changes.

If agents can discover suppliers, compare offers, negotiate within bounded authority, verify evidence, route payment, and preserve transaction state, some activities that once required managerial hierarchy may move outward into protocols, markets, or temporary collaborations. Other activities may move inward because they require tightly integrated context, specialized authority, risk management, or a shared body of institutional knowledge.

The coded enterprise does not necessarily produce either the enormous corporation or the tiny company. It makes the boundary more fluid.

The firm also cannot be reduced to transaction-cost avoidance. Kogut and Zander argued that firms function as social communities capable of creating, transferring, and recombining knowledge. Their advantage may reside not in a detachable instruction but in the organization’s combinative capacity: the way people with shared language and history bring distinct pieces of knowledge together.

This is one of the strongest objections to a simplistic business-as-code thesis. Some knowledge is tacit. Some is embodied. Some is distributed among participants who cannot fully state their own contribution. Some exists only in a shared practice. Some depends upon knowing when the recorded rule is irrelevant. Encoding the visible output does not necessarily capture the system that produced it.

But this objection does not halt the movement toward business as code. It expands its target. Where isolated knowledge cannot be encoded, competitive pressure moves outward to encode the conditions that make the knowledge usable.

The model needs context. Context needs provenance. Action needs authority. Authority needs a boundary. The boundary needs a definition of who is affected. The effect needs consent. The consequence needs witness. Failure needs recourse.

AI does not reveal that businesses are unnecessary. It reveals how much hidden code businesses contain.

When the Business Becomes Code

A company does not become code merely because it automates payroll, installs an enterprise system, or lets an agent answer customer questions. It becomes code when its operative institutional state is sufficiently legible for humans and machines to coordinate through it directly.

That state includes identity: what continuing entity is participating. It includes delegation: on whose behalf the participant may act. It includes authority: which decisions it can legitimately initiate, approve, or bind. It includes context: which records, definitions, policies, agreements, and circumstances govern. It includes consent: which effects the affected participants have permitted. It includes witness: what happened, under which state, and with what result. It includes memory, boundaries, exceptions, and the available paths of recourse after error, disagreement, breach, or harm.

Conventional enterprise software contains workflows. Business as code contains the constitutional state within which workflows become legitimate.

This does not require every institutional rule to become rigid or every ambiguity to disappear. A mature system must preserve the existence of ambiguity. It must know when the available representation is insufficient and when an action should stop, narrow, or return to a competent human or outer governance loop.

The goal is not to convert judgment into inflexible policy. It is to keep judgment from becoming an invisible source of unreviewable power.

Business as code is therefore not the company without people. It is the company whose identity, permissions, actions, dependencies, and consequences no longer rely entirely on people carrying them implicitly. People still participate. They decide, care, create, refuse, negotiate, interpret, repair, and change the system. But they no longer need to be the sole substrate through which the organization remembers what it is.

The Clean Start

There is an ethical defect at the foundation of the present transition.

The earliest general-purpose models were made possible by converting vast quantities of human expression into privately controlled machine capability without a meaningful, provenance-bearing chain of consent. Availability was treated as permission.

It was not.

Dataset-governance research has repeatedly emphasized that the origin, composition, collection context, intended use, and limitations of training data materially affect the systems produced from it. Documentation frameworks such as Datasheets for Datasets exist precisely because data cannot responsibly be treated as context-free raw material whose mere availability resolves the ethical and technical questions surrounding its use.

We cannot practically unscramble the first generation of foundation models. Meaningful consent cannot be manufactured retroactively. But neither must society accept permanent private enclosure as the only possible consequence.

A principled settlement would be that models substantially produced through broad, unconsented capture of public human expression should become free public infrastructure. This does not mean proprietary artificial intelligence must end. It means the inherited baseline should not remain a private tollbooth.

The historical capability becomes a commons. From that shared starting point, proprietary advantage may arise from genuinely new contribution, specialized systems, consented data, trusted relationships, attributable participation, and auditable provenance.

This is the civilizational-amnesty proposal developed in Changing the Optimization Primitive: do not pretend the extraction was consentful, and do not attempt the impossible task of reversing every historical training event. Socialize the inherited capability, then change the future default so that commercial optimization must be able to demonstrate a consent path.

Openness is not absolution. Opening a model does not restore consent, compensate every contributor, remove embedded harm, or answer every question of provenance. It prevents one additional harm: the conversion of a collective, unconsented inheritance into a permanent private head start.

A collective inheritance should not become a permanent private tollbooth merely because its first enclosure was technically difficult.

The clean start also accelerates the larger transition. Once baseline intelligence is broadly available, companies can no longer rely indefinitely on generic cognitive capability as their primary distinction. They must compete on what they can build around it—with permission.

When Knowledge Stops Being the Moat

No advantage based primarily on reproducible information or reproducible skill is permanently secure.

That does not mean every advantage disappears. Physical resources remain scarce. Location matters. Capital matters. Energy, infrastructure, manufacturing capacity, legal privilege, regulation, distribution, and network position matter. Some organizational capabilities are path-dependent and difficult to imitate. Some knowledge remains tacit or inseparable from communities of practice. The ability to sense change, reconfigure resources, and repeatedly create new combinations may outlast any single body of proprietary knowledge.

But the informational component of advantage becomes progressively vulnerable. A recipe that can be inferred becomes a commodity. A procedure that can be demonstrated becomes a candidate for automation. A professional pattern that can be evaluated across enough examples becomes a candidate for a model. A model available to one company becomes a product. A product becomes a service. A service becomes a protocol. A protocol becomes expected infrastructure.

The endpoint is not that nothing remains scarce. It is that information alone becomes progressively less able to explain the scarcity.

The strongest version of the thesis is therefore not that every form of knowledge will be extracted. It is that every reproducible capability will face continued attempts at extraction and codification, and successful codification will move competitive advantage to another layer.

Eventually, that movement reaches the boundary between information and participation.

A competitor can copy the answer. It cannot automatically copy the conditions under which someone is willing to tell you the real question.

Relationship Is Productive Infrastructure

“What remains will be relationships” can sound sentimental, like advice to improve customer service after the serious economic machinery has finished operating.

That is not the claim. Relationship is part of the machinery.

A business relationship is accumulated mutual state: prior conduct, contextual knowledge, expectations, permissions, reputation, dependency, unresolved obligation, shared risk, demonstrated repair, and confidence about future interaction. A record of that state can be copied. The relationship cannot be copied without the continuing participation of the people or institutions involved.

Research in strategy has long recognized that valuable resources may reside between firms rather than entirely within them. Dyer and Singh’s relational view identifies relation-specific assets, knowledge-sharing routines, complementary capabilities, and effective governance as possible sources of interorganizational competitive advantage. Randomized research on interfirm networking has also found that structured business relationships can improve revenue, profit, management practices, borrowing, and partnership formation.

Relationships are productive because they change what participants can safely and efficiently do together. A trusted supplier may disclose a risk before it becomes visible. An employee may reveal an error before it becomes a crisis. A customer may share an incomplete need rather than presenting a polished procurement specification. A community may provide legitimacy that cannot be purchased through advertising. A partner may tolerate ambiguity while a joint possibility is still forming. A creditor may permit flexibility during disruption. A participant may remain present after breach because prior conduct supports a credible expectation of repair.

These are not soft benefits surrounding the transaction. They are conditions that make possible transactions which could not be fully specified, priced, or enforced in advance.

Formal code does not necessarily destroy this relational capacity. Formal agreements and relational governance can function as complements. Clarity can reduce ambiguity, organize expectations, and support trust rather than replace it.

The right distinction is therefore not code or relationship. It is code that supports voluntary relationship versus code that captures and controls it.

The conditions, boundaries, permissions, memories, and obligations of a relationship can become more legible through code. Code cannot manufacture the other party’s continuing willingness to participate.

A competitor can copy your answer. It cannot copy why someone trusts you with the real question.

The Durable Moat Is Relational Capability

Relationship alone is not a moat. A company can inherit relationships, purchase access to them, trap people inside them, or preserve them long after they have ceased to be beneficial.

The durable capability is more demanding: the ability to form, govern, deserve, repair, and sustain voluntary coordination.

This is relational capability. It includes the ability to earn truthful disclosure, carry context without exploiting it, make authority understandable, coordinate without requiring domination, preserve confidence during uncertainty, recognize and repair breach, allow boundaries to change, and remain worth choosing when exit is possible.

That final condition matters. Captivity also produces continuity.

A relationship is not cooperative merely because it lasts. Lock-in can look like loyalty from the perspective of the captor. Silence can look like satisfaction to an institution that has made complaint dangerous. Compliance can look like consent where exit would destroy a person’s income, identity, standing, or access to unrelated parts of life.

The final moat cannot be possession of relationships. It is the ability to deserve continued participation in them.

Predation Is an Information-Degrading Strategy

Predation can be locally successful. A company may increase revenue by concealing costs, exploiting dependency, extracting uncompensated contribution, narrowing exit, controlling information, surveilling behavior, or converting asymmetry into leverage. Such strategies may outperform cooperation in the transaction directly in front of them.

But participants adapt. Where disclosure creates vulnerability, people disclose less. Where honesty is punished, people become performative. Where metrics become weapons, people optimize the metric. Where every contribution is captured, people stop contributing their best thought in the shared space. Where the system cannot distinguish refusal from disobedience, participants conceal their boundaries until they can escape.

An extractive institution may collect more data while receiving less truth.

This matters more as generic machine capability becomes abundant. When many firms can access broadly comparable models, their relative advantage depends increasingly on the quality of the context those models receive. High-quality context includes information that cannot reliably be compelled: uncertainty, local knowledge, weak signals, emerging danger, dissatisfaction, informal exceptions, fear, intention, partial ideas, and evidence of failure.

People offer this kind of information most cleanly when they believe disclosure will not be used predictably against them.

Cooperation is therefore not merely kinder. Under many conditions, it is a higher-fidelity information strategy.

Predation optimizes the exchange in front of it. Cooperation preserves the source of future exchange.

This is how the dissolution of informational moats can become good for humanity. When generic capability is scarce, advantage can be maintained through enclosure. When generic capability becomes common, differentiation moves toward the ability to receive better context, coordinate more faithfully, and sustain participation over time.

Predation remains possible, but it increasingly consumes the relational substrate on which high-quality intelligence depends.

Cooperation Does Not Happen by Itself

Technological abundance does not automatically move humanity from predation to cooperation. Research on cooperation argues almost the opposite: cooperation persists when supporting mechanisms exist.

Repeated interaction can make reciprocal cooperation viable. Reputation can carry the consequence of one interaction into another. Network structure can protect or undermine cooperative behavior. Monitoring and proportionate sanctions can discourage free riding. Meaningful participation in rule formation can improve the durability of shared governance.

Cooperation requires architecture.

In the coded enterprise, that architecture includes identity sufficient for accountable continuity, boundaries defining who and what is participating, legible delegation, consent that can change over time, memory of relevant prior conduct, witness of consequential transitions, recourse after breach, and meaningful exit.

The objective is not to eliminate trust. It is to stop requiring trust to survive avoidable ambiguity.

Code can preserve what was agreed, which definitions governed, who held authority, what evidence existed, where permission ended, what changed, and which repair path remains open. This allows cooperation to scale beyond the limits of personal familiarity without pretending that a record is the same thing as a relationship.

Cooperation is not the absence of structure. It is structure that participants can enter, understand, influence, witness, and leave.

The Rival End-State

The conversion toward business as code may be inevitable. Its constitution is not.

The same technologies could produce an economy of concentrated model ownership, pervasive behavioral surveillance, opaque algorithmic management, automatic exclusion, non-negotiable machine-generated terms, and agents capable of binding people to consequences they did not meaningfully understand.

Algorithmic management already demonstrates that software can centralize direction, evaluation, discipline, and control while making the reasoning behind managerial action more difficult to inspect or contest. Digital transformation does not inherently decentralize authority. It can intensify it.

Technological advantage may also concentrate markets rather than distribute them. Research on superstar firms associates changing technologies and market conditions with increased concentration and declining labor shares in many industries.

There are therefore at least two forms of business as code. One treats identity as surveillance; the other treats it as accountable continuity. One treats delegation as impersonation; the other makes authority explicit and revocable. One reduces consent to a historical click; the other treats consent as continuing permission. One treats memory as capture; the other treats it as governed shared state. One uses automation to remove contestability; the other uses automation to make consequence legible. One turns relationship into lock-in; the other protects voluntary continuation and exit.

The conversion is inevitable. The constitution is not.

Business as code can make exploitation frictionless, or it can make consequence impossible to hide.

The Constitution of the Coded Business

A coded business should not be judged only by whether it achieves its objective. It should also be judged by whether its authority, permissions, effects, and consequences remain legitimate and legible while it does so.

Identity must be persistent enough for accountability but proportionate enough to avoid unnecessary exposure. Delegation must be bounded: an agent should carry revocable authority, not unrestricted access to the person or institution it represents. Context must be anchored so that agreements, records, definitions, and policy versions can be identified without depending on somebody’s memory of what the system probably meant.

Consent must govern effects. The fact that a system can perform an action does not mean it has permission to affect everyone touched by that action. Witness must preserve consequence. Important transitions must leave enough evidence to reconstruct what occurred, under what authority, and within which state.

Exceptions must remain possible because no formal representation will capture every relevant human condition. Recourse must survive automation because a decision without a reachable repair path is not merely efficient; it is potentially automated domination. Exit must be meaningful. A participant must be able to leave without the system destroying their identity, history, standing, or access to unrelated relationships.

Authority must also degrade safely. When an automated process can no longer preserve human safety, agency, standing, consent, or future possibility, it must stop, narrow itself, or reroute into a competent outer loop.

The coded enterprise needs more than an operating system. It needs a constitution.

The Big Hook Revealed

The first hook was convenience. People began using AI because it helped them do things.

The second hook was institutional dependence. Businesses discovered that work increasingly moved through models whether or not the company had designed a coherent architecture for it.

The third hook is interoperability. When customers, employees, partners, suppliers, regulators, and institutions begin arriving through agents, a business must expose enough of itself to be recognized, questioned, negotiated with, trusted, and held to account.

Every serious organization then needs access to a connective layer carrying identity, context, delegation, authority, consent, provenance, witness, memory, boundaries, and recourse.

That layer is the Big Hook.

It would be dangerous for it to belong entirely to one model provider, one identity company, one payment network, one enterprise platform, or one state. The cooperative alternative is an interoperable substrate: a common grammar through which distinct people, agents, firms, and institutions can coordinate without surrendering unconditional control to the party operating the infrastructure.

Whatever can be made legible will be subjected to encoding. Whatever can be encoded will be subjected to replication. Whatever can be replicated will lose some of its scarcity. As capability becomes common, coordination becomes decisive.

Coordination depends on the willingness of distinct actors to continue together. That willingness cannot be extracted indefinitely. It must be deserved.

Business will not disappear when business becomes code. It will become more visible. Its promises will appear as state. Its permissions will appear as boundaries. Its authority will appear as delegation. Its failures will appear as witnessed consequence.

Its relationships will no longer be treated as soft material surrounding the real machinery of the firm. They will be recognized as the machinery through which everything else remains possible.

Knowledge will still matter. Skill will still matter. Judgment will still matter. But none will remain scarce merely because it once lived inside a person, department, profession, or closed system. Whatever can be outed will continue its movement toward infrastructure.

What cannot be copied so easily is the willingness of another being to remain in relationship with you.

When every business becomes code, relationship becomes the source code.

References